Which description best matches reinforcement learning?

• A. Algorithm uses input data without output data
• B. Algorithm learns to perform a task to try and maximise reward
• C. Algorithm requires supervised labels to learn
• D. Algorithm minimizes error on labeled pairs

Answer: (B)

Reinforcement learning is about an agent learning what actions to take in an environment to maximize cumulative rewards over time. The agent acts, observes the result (next state) and the reward, and gradually improves its policy by balancing trying new actions (exploration) with using known good actions (exploitation). This feedback loop is what drives the agent to achieve higher long-term returns.

This makes the description that focuses on learning to perform a task to maximize reward the best match. It captures the core idea of acting to gain rewards and improve over time through interaction.

The other descriptions describe different learning paradigms. Using input data without output data points to unsupervised learning, which doesn’t rely on a reward signal. Requiring supervised labels or minimizing error on labeled pairs aligns with supervised learning, where correct outputs are provided for training data. Reinforcement learning relies on reward signals rather than explicit correct labels.